Vellum Assistant
A personal assistant that retains context, follows up proactively, and acts across your connected channels.
Per-dimension scores and reasoning
Evidence shows sandboxing, permission gating, credential separation, but implementation details not shown in provided files, so some criteria get 2. Deductions: data flow transparency only mentioned, not detailed; rollback not clear; source attribution only mentioned, not deep.
Self-consistency: README and SECURITY.md consistent, but no code verification. Dependency availability: package.json and workflows present, but not verified. Failure messages: no specific error handling examples. Deductions: lack of runtime evidence.
Audience and scenarios: README describes various use cases, but no concrete examples. Capability boundaries: security model doc exists, but not detailed. Trigger precision: no specific trigger mechanisms. Environment fit: multi-platform support, but not verified. Deductions: static review cannot verify actual adaptation.
Information architecture: documentation structure exists, but not complete. Install notes: CLI install steps present, but not verified. Naming stability: no version history. Examples and FAQ: quick demo present, but no FAQ. Known limitations: not explicit. License: MIT, complete. Versioning and changelog: not provided. Maintenance responsibility: contributing guide exists, but not verified. Deductions: some docs missing.
Output usability: CLI and clients present, but not verified. Marginal value: unique features, but not verified. Cost-benefit: managed and self-hosted options, but not verified. Deductions: static review cannot assess actual effectiveness.
Claim traceability: some claims have doc support, but no evidence provided. Cross-source corroboration: multiple docs, but not cross-verified. Fact-inference separation: not explicit. Deductions: lack of independent verification.
- Static review cannot verify actual runtime behavior; all security claims need dynamic testing.
- Dependency security is based only on overrides and patchedDependencies in package.json; effectiveness of fixes not verified.
- Rollback mechanism not clear; confirm data recovery plan before deployment.
What does this agent do, and when should you use it?
Vellum Assistant is a personal AI assistant available as a Vellum-managed runtime or a locally hosted deployment. It extracts structured identity, preference, project, and event items from conversations, with source attribution, deduplication, and user- and channel-level isolation. Its behavior is stored in SOUL.md, its active focus and threads in NOW.md, and it keeps per-user reflection journals. One assistant and memory can be reached through macOS, iOS, web, voice, email, Telegram, Slack, and Twilio. In a sandbox and with permission for access to the actual machine, it can read and edit files, run commands, drive a browser, and periodically notify users about unfinished or upcoming work.
The assistant extracts structured identity, preference, project, and event records from conversation, then maintains eight memory types through local-default ONNX embeddings and hybrid dense-plus-sparse retrieval. It writes behavioral material to SOUL.md, keeps current focus and active threads in NOW.md, and rereads its notes hourly to identify unfinished or due-soon items before messaging the appropriate channel. With approval, it performs file reads and edits, command execution, and browser control in its own sandbox; unknown actors cannot read memory, trigger tools, or escalate. The CLI exposes vellum wake to start services, vellum client for terminal interaction, vellum ps to inspect running assistants, and vellum terminal to open a shell in a managed assistant container.
- A person who starts a request in Slack and wants to continue it in Telegram without losing the assistant's context.
- A knowledge worker who wants conversation-derived preferences, projects, and upcoming events retained and proactively surfaced when attention is needed.
- A macOS user who wants to approve an assistant to edit local files, run shell commands, or operate a browser.
- A self-hosting user who wants a locally run assistant reachable from a phone, tablet, or another computer.
- A team member who needs OAuth connections to services such as Slack, Notion, Google, HubSpot, Linear, Discord, Twitter, Telegram, or Twilio.
What are this agent's strengths and limitations?
- Its eight memory types have individual staleness windows and pair hybrid retrieval with attribution, deduplication, and per-user/per-channel isolation rather than relying on a single maintained notes file.
- It includes an hourly review loop for unfinished or due-soon work and states that notifications are routed to the right channel without interrupting an active conversation.
- Its documented security model centralizes actor trust, keeps credentials in a separate process, sandboxes tool calls, and defaults to deny.
- The same codebase supports managed and self-hosted operation while documenting paths for Anthropic, OpenAI, Gemini, Ollama, and other providers.
- The README identifies the desktop app as the primary focus and positions the CLI mainly for advanced users, contributors, and non-macOS environments.
- The supplied material does not document local deployment credentials, environment variables, persistent storage, or a full configuration path.
- File, command, and browser access is permission-gated, while unknown actors cannot access memory or tools; trust controls may constrain unattended workflows.
- Several channels and OAuth connections depend on external services, but the supplied material does not describe failure handling, rate limits, or migration procedures.
How do you install or deploy this agent?
The CLI is positioned for advanced users, contributors, and non-macOS environments. After installing Bun, run:
bun install -g vellum
vellum hatchTo install from source, run:
git clone https://github.com/vellum-ai/vellum-assistant.git
cd vellum-assistant
./setup.sh
source ~/.bashrc
vellum hatchManaged mode requires signing in through Vellum Cloud. The supplied README does not specify local-model credentials, environment variables, or the complete first-run configuration for local mode.
How do you use this agent?
After vellum hatch, start services with vellum wake and interact in a terminal through vellum client. Use vellum sleep to stop services while keeping data, and vellum ps to view running assistants. For multiple assistants, pass the assistant ID as the second command argument. When the assistant needs access to the actual machine, approve it using the documented one-time, ten-minute, or always permission options.
How does this agent compare with similar options?
The README positions Vellum as reducing the setup effort associated with a personal AI on OpenClaw, Hermes Agent, or Claude Code. It does not provide feature or performance benchmarks against those alternatives.